Statistical models for determining the optimal combination of biomarkers and their application in classification of medical data

Abstract

Introduction Area under a ROC curve (AUC) is a common criterion to assess the overall classification performance of the markers. In practice due to limited classification ability of a single marker, we are interested in combining markers linearly or nonlinearly to improve classification performance. Ramp AUC (RAUC) is a new statistical AUC-based method which can find such optimal combinations of markers. In this study, RAUC was used to find the optimal combinations of care indicators related to functional limitation as a complication of diabetes and accurately discriminate this outcome based on its underlying markers. Methods This cross-sectional study was conducted on 378 diabetic patients referred to diabetic centers of Ardebil and Tabriz during 2014–15. To have an accurate classification of diabetic patients according to their functional limitation status, RAUC method with RBF kernel was employed to look for optimal combination of care indicators. Classification performance of the model was evaluated by AUC and compared with logistic regression, support vector machine (SVM) and generalized additive model (GAM) via training and test validation method. Results Out of 378 diabetics, 67,46% had functional limitation. RAUC had a test dataset AUC equal 1 and outperformed logistic (AUC=.79), GAM (AUC=.82), SVM with linear kernel (AUC=.67) and was slightly better than SVM with RBF kernel (AUC=.98). Conclusion110 There was strong nonlinearity in data and RAUC with RBF kernel which is a nonlinear combination of markers, could detect this pattern.